SS was performed using a custom-built system at AFRL—colloquially referred to as LEROY—that is capable of fully automated multi-step materialographic sample preparation and multi-modal microscopy workflows, including the collection of scanning electron microscopy data [10, 11]. Materialographic sample preparation and cleaning processes within the LEROY system are performed by a significantly modified, second-generation RoboMet.3D™ outfitted with 300 mm dia. polishing pads, which can achieve a wide range of SS thickness values while potentially finishing with nanometer-scale abrasives to create low-damage surfaces that are sufficient for EBSD-based mapping. The RoboMet.3D™ within the LEROY system includes a Zeiss AxioVert inverted optical microscope that can be used for bright field, dark field and polarized light microscopy collection. The LEROY system also incorporates a Tescan Vega 3 XMH LaB6 SEM with automated load lock exchange and Bruker Quantax e-Flash 1000 EBSD detector, and two additional Zeiss optical microscopes (Zeiss AxioImager.Z2 and AxioObserver), allowing the user to select and optimize the most appropriate modes of microscopy to include in the SS experiment for subsequent analysis.
The mechanical polishing workflow of the experiment consisted of two material removal steps that were performed by the RoboMet.3D™, using a constant force of ~ 24.5 N. The primary material removal step utilized 1 μm polycrystalline diamond abrasive in a water-based suspension (Allied High Tech) diluted with a concurrent drip of distilled water on a woven acetate polishing pad (DAC, Struers) for approximately 5 min at a rotational speed of 100 rpm. The second polishing step utilized 40 nm-scale colloidal silica suspension (non-stick rinsable, Allied High Tech.) diluted on a non-woven porous polyurethane polishing pad (Final A, Allied High Tech.) with a concurrent drip of distilled water for 2 min at a rotational speed of 100 rpm. After each of these two steps, a long-nap cloth (Final Pol, Allied High Tech.) with flowing tap water was used to clean the sample surface of polishing media (30 s at a rotational speed of 75 rpm) to minimize grit cross-contamination. Ultrasonic cleaning using pure Ethanol as the immersion fluid was performed prior to microscopic data collection, using a sequence of multiple beakers and nitrogen gas drying to minimize contaminants on the freshly exposed SS surface. Note that cube-corner microhardness indents were intermittently introduced to the SS surface (via MTS NanoXP Nanoindentation system, 9 N indentation load) throughout the experiment to the titanium support structure that surrounded the tensile sample, as a contingency protocol to quantify SS material removal rates [12], but were not used as optical microscope focus value measurements were deemed sufficient.
The microscopy workflow consisted of three types of data collection: bright-field optical montage images, EBSD maps, and backscatter SEM images, which are described in the remainder of this section. Bright-field optical images were included for two purposes. First, these were used as a data mode that allowed for robust registration of the volumetric stack of serial section microscopy data. Systematic errors in section-to-section alignment can occur when only one dominant feature is present in each section, e.g., the tensile sample cross-section, and a metric that maximizes the overlap of this dominant feature is used for stack registration [13]. This was of particular concern as the radiographic imaging of the metallographic mount indicated that the longitudinal axis of the tensile sample was slightly misaligned relative to the SS surface normal. To avoid this problem, the use of materialographic mount material that contained carbon filler particles, combined with the use of large-area montage image data collection that created images with thousands of randomly oriented particles, enabled the use of the SIFT algorithm [14] to calculate the optimal registration. Second, a common method to measure the depth of the SS surface is by recording the value of the optical microscope focus motor position [9], which can be effective when both high numerical aperture objective lenses (which have a narrow depth of field) are used in conjunction with automated focusing algorithms.
A Zeiss AxioVert 200 M inverted optical microscope outfitted with a AxioCam MRc5 camera was used to collect a bright-field 12-bit grayscale images from the SS surface, using a 20× objective lens (Epiplan, 0.4 NA) and a 1.0× optivar, resulting in an image pixel dimension of 0.52 µm. For the SS Region-Of-Interest (ROI), a 144 tile (12 × 12) montage using a single focus point and ~ 50% image overlap was collected at each section, which covered an area of approximately 6.2 × 4.6 mm. This relatively large ROI (compared to the tensile sample cross section) also included subregions of the titanium support structure with microhardness indents for contingency stack registration and depth measurements. A highly cropped region of the optical montage image data (highlighting just the tensile sample cross section) is shown in Fig. 4a. Three additional 3 × 3 montages were collected—one at the center of the ROI and two at opposing ends of the sample within the offset ledge—using the same microscope conditions but with autofocusing at every third tile, to allow for differential SS depth measurements.
Electron Backscatter Diffraction maps were a key data mode for this experiment in order to provide a spatially resolved micron-scale characterization of the local crystal orientation. A single 600 × 600 pixel map with 1 µm pixel dimensions was collected at each section, as shown in Fig. 4b. These maps were collected at 70° stage tilt, using a 20 kV accelerating voltage and a high (5–13 nA) probe current, and 8X binning (80 × 60 EBSD patterns), enabling the Bruker detector to operate at 550 points per second, resulting in map collection times of approximately 11 min. Note that while the individual EBSD patterns from the tensile sample contained clearly defined Kikuchi bands, there was often a few percent of pixels that were not initially indexed by the Bruker Esprit 1.9.4 software, therefore both the original EBSD patterns (*.bcf) and the indexed data (*.ctf) were saved at each section to enable dictionary-based reindexing analysis [15]. Saving all of the EBSD patterns significantly increased the total amount of disk storage space for the experiment, as each *.bcf file was approximately 2 GB.
It is known that geometric distortions can commonly occur in EBSD data scans, which are accentuated by the large tilt of the sample surface relative to the incident electron beam [16]. As will be discussed in the following section, a data fusion approach was employed to address both geometrical distortions and stack registration by a non-linear spatial transform of the raw EBSD data to the registered stack of optical microscopy data. Initial attempts to register the EBSD maps directly to the optical image failed, as there were few common features in both data modes. Therefore, backscattered electron (BSE) images at normal incidence (0° stage tilt) were also collected at each section, as an intermediary data type to facilitate EBSD stack registration. As shown in Fig. 4c, the BSE images contain significant electron channeling contrast that correlated to the crystalline grain structure, and this contrast feature was used to register EBSD maps to the corresponding BSE images. Moreover, numerous micrometer-scale precipitates and pores within the tensile sample were also observable in both the optical and BSE images, which enabled registration of these data modes. Occasional and random local intensity fluctuations of the BSE signal were observed during image collection, thus five 1024 × 1024 pixel images with 0.59 µm pixel dimensions with 16-bit grayscale depth were collected at each section, using an accelerating voltage of 20 kV and a probe current of ~ 1 nA, and resulted in a collection time of approximately 3 min. These images were fused post-collection into a single image that eliminated the intensity fluctuations, using Fiji [17] to perform SIFT registration of the stack of 5 images followed by a z-projection median filter.
Considering all of the aforementioned steps of the SS experiment, the total cycle time for one section was approximately 63 min, which includes additional activities such as robotic manipulation within and between the various subsystems (e.g., polishing pad exchanges for RoboMet.3D, exchange of the sample into and out of the SEM via a load lock), and real-time optimization of key SEM acquisition parameters (e.g., working distance, focus, and brightness/contrast). The cumulative experiment consisted of 1000 sections that completely encompassed the volume denoted by the two outer gold markers, although only a subset of the 1000 sections was required for the Challenge 4 data. The initial 900 sections were collected in a 49-day span before a SEM vacuum pump failure occurred, which resulted in a few week delay before the final 100 sections were collected.
Figure 5a shows a plot of the calculated depth of the SS surface from optical microscope focus values as a function of section number for the experiment, and an ordinary least squares linear regression fit to this data. These material removal data were generated from averaging the microscope focus values of the aforementioned “additional 3 × 3 montages” that contain 3 autofocus operations per montage. The residuals for the linear regression fit are shown in Fig. 5b, and the authors note that occasional outlier values are present in the data, which were included in the calculation of the linear fit. For 3D reconstruction of the SS data, the authors selected to use a constant value for the voxel depth of 0.97 µm per section for the first 900 sections of the data set, and this selection is supported by the coefficient of determination (R2) value of 0.9998 for the linear fit. The choice of a constant depth value also greatly simplified some aspects of the reconstruction workflow. The residual errors to the linear fit in Fig. 5b are not entirely randomly distributed, and therefore efforts to calculate non-uniform (section-to-section) depth values will likely enhance the accuracy of subsequent 3D reconstructions, but have not been performed to date.